Breast cancer is one of the leading causes of cancer-related mortality among women worldwide, where early and accurate diagnosis plays a decisive role in improving survival outcomes. Breast ultrasound imaging is widely adopted in clinical practice due to its nonionizing nature, real-time capability, cost-effectiveness, and suitability for dense breast tissue. However, ultrasound images are inherently affected by speckle noise, low contrast, heterogeneous tissue appearance, and operator dependency, which significantly limit diagnostic reliability. To address these challenges, this doctoral thesis proposes a robust, automated computer-aided diagnosis (CAD) framework for breast tumor analysis using deep learning and machine learning techniques. The proposed framework follows a threestage pipeline comprising lesion segmentation, deep learning-based feature extraction, and tumor classification. In the first stage, image preprocessing was performed using anisotropic diffusion filtering to suppress speckle noise while preserving lesion boundaries. Breast lesions are then segmented using an advanced U-Net++ architecture. Experimental evaluation on the BUSI dataset demonstrates that U-Net++ achieves a Dice similarity coefficient of 0.911 and a mean Intersection-over-Union (IoU) of 0.838, outperforming standard U-Net and Efficient U-Net models and providing superior boundary localization. In the second stage, deep learning-based feature extraction is performed using a lightweight MobileNetV2 architecture, selected after comparative evaluation with InceptionV3, EfficientNetB7, ResNet50, and VGG16. MobileNetV2 offers an optimal balance between discriminative feature representation and computational efficiency, making it suitable for real-time and resource-constrained clinical environments. In the final stage, multiple machine learning classifiers, including Artificial Neural Networks (ANN), Support Vector Machines (SVM), Random Forest (RF), K-Nearest Neighbors (KNN), and Long Short-Term Memory (LSTM) networks, are evaluated for benign, malignant, and normal tumor classification. Ablation analysis confirms the contribution of each pipeline component, with accuracy dropping to 91.2% without feature extraction and 88.70% without segmentation. The fully optimized framework achieves an overall classification accuracy of 96.58%, highlighting the cumulative benefit of preprocessing, segmentation, and deep feature learning.Furthermore, the thesis explores the use of a conditional Generative Adversarial Network (cGAN) for dataset augmentation. The GAN-enhanced classification model achieves an average cross-validation accuracy of 91.4% ± 1.2%, with malignant tumor recall reaching 93.1%, and demonstrates strong generalization on a separate real-image test set with 91.03% accuracy. Overall, the proposed CAD framework significantly reduces operator dependency, improves diagnostic robustness, and demonstrates strong potential for realtime clinical deployment. This work contributes a scalable, lightweight, and highperformance solution for breast ultrasound-based cancer diagnosis and provides a foundation for future extensions involving end-to-end learning, multimodal imaging, and clinically validated deployment.